tjh.dev
Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
cpuidle: add haltpoll governor
The cpuidle_haltpoll governor, in conjunction with the haltpoll cpuidle
driver, allows guest vcpus to poll for a specified amount of time before
halting.
This provides the following benefits to host side polling:
1) The POLL flag is set while polling is performed, which allows
a remote vCPU to avoid sending an IPI (and the associated
cost of handling the IPI) when performing a wakeup.
2) The VM-exit cost can be avoided.
The downside of guest side polling is that polling is performed
even with other runnable tasks in the host.
Results comparing halt_poll_ns and server/client application
where a small packet is ping-ponged:
host --> 31.33
halt_poll_ns=300000 / no guest busy spin --> 33.40 (93.8%)
halt_poll_ns=0 / guest_halt_poll_ns=300000 --> 32.73 (95.7%)
For the SAP HANA benchmarks (where idle_spin is a parameter
of the previous version of the patch, results should be the
same):
hpns == halt_poll_ns
idle_spin=0/ idle_spin=800/ idle_spin=0/
hpns=200000 hpns=0 hpns=800000
DeleteC06T03 (100 thread) 1.76 1.71 (-3%) 1.78 (+1%)
InsertC16T02 (100 thread) 2.14 2.07 (-3%) 2.18 (+1.8%)
DeleteC00T01 (1 thread) 1.34 1.28 (-4.5%) 1.29 (-3.7%)
UpdateC00T03 (1 thread) 4.72 4.18 (-12%) 4.53 (-5%)
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
authored by
Marcelo Tosatti
and committed by
Rafael J. Wysocki
2cffe9f6
7d4daeed
+240
+78
Documentation/virtual/guest-halt-polling.txt
+78
Documentation/virtual/guest-halt-polling.txt
···
1
+
Guest halt polling
2
+
==================
3
+
4
+
The cpuidle_haltpoll driver, with the haltpoll governor, allows
5
+
the guest vcpus to poll for a specified amount of time before
6
+
halting.
7
+
This provides the following benefits to host side polling:
8
+
9
+
1) The POLL flag is set while polling is performed, which allows
10
+
a remote vCPU to avoid sending an IPI (and the associated
11
+
cost of handling the IPI) when performing a wakeup.
12
+
13
+
2) The VM-exit cost can be avoided.
14
+
15
+
The downside of guest side polling is that polling is performed
16
+
even with other runnable tasks in the host.
17
+
18
+
The basic logic as follows: A global value, guest_halt_poll_ns,
19
+
is configured by the user, indicating the maximum amount of
20
+
time polling is allowed. This value is fixed.
21
+
22
+
Each vcpu has an adjustable guest_halt_poll_ns
23
+
("per-cpu guest_halt_poll_ns"), which is adjusted by the algorithm
24
+
in response to events (explained below).
25
+
26
+
Module Parameters
27
+
=================
28
+
29
+
The haltpoll governor has 5 tunable module parameters:
30
+
31
+
1) guest_halt_poll_ns:
32
+
Maximum amount of time, in nanoseconds, that polling is
33
+
performed before halting.
34
+
35
+
Default: 200000
36
+
37
+
2) guest_halt_poll_shrink:
38
+
Division factor used to shrink per-cpu guest_halt_poll_ns when
39
+
wakeup event occurs after the global guest_halt_poll_ns.
40
+
41
+
Default: 2
42
+
43
+
3) guest_halt_poll_grow:
44
+
Multiplication factor used to grow per-cpu guest_halt_poll_ns
45
+
when event occurs after per-cpu guest_halt_poll_ns
46
+
but before global guest_halt_poll_ns.
47
+
48
+
Default: 2
49
+
50
+
4) guest_halt_poll_grow_start:
51
+
The per-cpu guest_halt_poll_ns eventually reaches zero
52
+
in case of an idle system. This value sets the initial
53
+
per-cpu guest_halt_poll_ns when growing. This can
54
+
be increased from 10000, to avoid misses during the initial
55
+
growth stage:
56
+
57
+
10k, 20k, 40k, ... (example assumes guest_halt_poll_grow=2).
58
+
59
+
Default: 50000
60
+
61
+
5) guest_halt_poll_allow_shrink:
62
+
63
+
Bool parameter which allows shrinking. Set to N
64
+
to avoid it (per-cpu guest_halt_poll_ns will remain
65
+
high once achieves global guest_halt_poll_ns value).
66
+
67
+
Default: Y
68
+
69
+
The module parameters can be set from the debugfs files in:
70
+
71
+
/sys/module/haltpoll/parameters/
72
+
73
+
Further Notes
74
+
=============
75
+
76
+
- Care should be taken when setting the guest_halt_poll_ns parameter as a
77
+
large value has the potential to drive the cpu usage to 100% on a machine which
78
+
would be almost entirely idle otherwise.
+11
drivers/cpuidle/Kconfig
+11
drivers/cpuidle/Kconfig
···
33
33
Some workloads benefit from using it and it generally should be safe
34
34
to use. Say Y here if you are not happy with the alternatives.
35
35
36
+
config CPU_IDLE_GOV_HALTPOLL
37
+
bool "Haltpoll governor (for virtualized systems)"
38
+
depends on KVM_GUEST
39
+
help
40
+
This governor implements haltpoll idle state selection, to be
41
+
used in conjunction with the haltpoll cpuidle driver, allowing
42
+
for polling for a certain amount of time before entering idle
43
+
state.
44
+
45
+
Some virtualized workloads benefit from using it.
46
+
36
47
config DT_IDLE_STATES
37
48
bool
38
49
+1
drivers/cpuidle/governors/Makefile
+1
drivers/cpuidle/governors/Makefile
+150
drivers/cpuidle/governors/haltpoll.c
+150
drivers/cpuidle/governors/haltpoll.c
···
1
+
// SPDX-License-Identifier: GPL-2.0
2
+
/*
3
+
* haltpoll.c - haltpoll idle governor
4
+
*
5
+
* Copyright 2019 Red Hat, Inc. and/or its affiliates.
6
+
*
7
+
* This work is licensed under the terms of the GNU GPL, version 2. See
8
+
* the COPYING file in the top-level directory.
9
+
*
10
+
* Authors: Marcelo Tosatti <mtosatti@redhat.com>
11
+
*/
12
+
13
+
#include <linux/kernel.h>
14
+
#include <linux/cpuidle.h>
15
+
#include <linux/time.h>
16
+
#include <linux/ktime.h>
17
+
#include <linux/hrtimer.h>
18
+
#include <linux/tick.h>
19
+
#include <linux/sched.h>
20
+
#include <linux/module.h>
21
+
#include <linux/kvm_para.h>
22
+
23
+
static unsigned int guest_halt_poll_ns __read_mostly = 200000;
24
+
module_param(guest_halt_poll_ns, uint, 0644);
25
+
26
+
/* division factor to shrink halt_poll_ns */
27
+
static unsigned int guest_halt_poll_shrink __read_mostly = 2;
28
+
module_param(guest_halt_poll_shrink, uint, 0644);
29
+
30
+
/* multiplication factor to grow per-cpu poll_limit_ns */
31
+
static unsigned int guest_halt_poll_grow __read_mostly = 2;
32
+
module_param(guest_halt_poll_grow, uint, 0644);
33
+
34
+
/* value in us to start growing per-cpu halt_poll_ns */
35
+
static unsigned int guest_halt_poll_grow_start __read_mostly = 50000;
36
+
module_param(guest_halt_poll_grow_start, uint, 0644);
37
+
38
+
/* allow shrinking guest halt poll */
39
+
static bool guest_halt_poll_allow_shrink __read_mostly = true;
40
+
module_param(guest_halt_poll_allow_shrink, bool, 0644);
41
+
42
+
/**
43
+
* haltpoll_select - selects the next idle state to enter
44
+
* @drv: cpuidle driver containing state data
45
+
* @dev: the CPU
46
+
* @stop_tick: indication on whether or not to stop the tick
47
+
*/
48
+
static int haltpoll_select(struct cpuidle_driver *drv,
49
+
struct cpuidle_device *dev,
50
+
bool *stop_tick)
51
+
{
52
+
int latency_req = cpuidle_governor_latency_req(dev->cpu);
53
+
54
+
if (!drv->state_count || latency_req == 0) {
55
+
*stop_tick = false;
56
+
return 0;
57
+
}
58
+
59
+
if (dev->poll_limit_ns == 0)
60
+
return 1;
61
+
62
+
/* Last state was poll? */
63
+
if (dev->last_state_idx == 0) {
64
+
/* Halt if no event occurred on poll window */
65
+
if (dev->poll_time_limit == true)
66
+
return 1;
67
+
68
+
*stop_tick = false;
69
+
/* Otherwise, poll again */
70
+
return 0;
71
+
}
72
+
73
+
*stop_tick = false;
74
+
/* Last state was halt: poll */
75
+
return 0;
76
+
}
77
+
78
+
static void adjust_poll_limit(struct cpuidle_device *dev, unsigned int block_us)
79
+
{
80
+
unsigned int val;
81
+
u64 block_ns = block_us*NSEC_PER_USEC;
82
+
83
+
/* Grow cpu_halt_poll_us if
84
+
* cpu_halt_poll_us < block_ns < guest_halt_poll_us
85
+
*/
86
+
if (block_ns > dev->poll_limit_ns && block_ns <= guest_halt_poll_ns) {
87
+
val = dev->poll_limit_ns * guest_halt_poll_grow;
88
+
89
+
if (val < guest_halt_poll_grow_start)
90
+
val = guest_halt_poll_grow_start;
91
+
if (val > guest_halt_poll_ns)
92
+
val = guest_halt_poll_ns;
93
+
94
+
dev->poll_limit_ns = val;
95
+
} else if (block_ns > guest_halt_poll_ns &&
96
+
guest_halt_poll_allow_shrink) {
97
+
unsigned int shrink = guest_halt_poll_shrink;
98
+
99
+
val = dev->poll_limit_ns;
100
+
if (shrink == 0)
101
+
val = 0;
102
+
else
103
+
val /= shrink;
104
+
dev->poll_limit_ns = val;
105
+
}
106
+
}
107
+
108
+
/**
109
+
* haltpoll_reflect - update variables and update poll time
110
+
* @dev: the CPU
111
+
* @index: the index of actual entered state
112
+
*/
113
+
static void haltpoll_reflect(struct cpuidle_device *dev, int index)
114
+
{
115
+
dev->last_state_idx = index;
116
+
117
+
if (index != 0)
118
+
adjust_poll_limit(dev, dev->last_residency);
119
+
}
120
+
121
+
/**
122
+
* haltpoll_enable_device - scans a CPU's states and does setup
123
+
* @drv: cpuidle driver
124
+
* @dev: the CPU
125
+
*/
126
+
static int haltpoll_enable_device(struct cpuidle_driver *drv,
127
+
struct cpuidle_device *dev)
128
+
{
129
+
dev->poll_limit_ns = 0;
130
+
131
+
return 0;
132
+
}
133
+
134
+
static struct cpuidle_governor haltpoll_governor = {
135
+
.name = "haltpoll",
136
+
.rating = 21,
137
+
.enable = haltpoll_enable_device,
138
+
.select = haltpoll_select,
139
+
.reflect = haltpoll_reflect,
140
+
};
141
+
142
+
static int __init init_haltpoll(void)
143
+
{
144
+
if (kvm_para_available())
145
+
return cpuidle_register_governor(&haltpoll_governor);
146
+
147
+
return 0;
148
+
}
149
+
150
+
postcore_initcall(init_haltpoll);